Will the Market Fix the Market? A Theory of Stock Exchange Competition and Innovation - Harvard University
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Will the Market Fix the Market?
A Theory of Stock Exchange Competition and Innovation∗
Eric Budish†, Robin S. Lee‡, and John J. Shim§
May 6, 2019
Abstract
As of early 2019, there are 13 stock exchanges in the U.S., across which over 1 trillion shares ($50
trillion) are traded annually. All 13 exchanges use the continuous limit order book market design, a design
that gives rise to latency arbitrage—arbitrage rents from symmetrically observed public information—
and the associated high-frequency trading arms race (Budish, Cramton and Shim, 2015). Will the
market adopt new market designs that address the negative aspects of high-frequency trading? This
paper builds a theoretical model of stock exchange competition to answer this question. Our model,
shaped by institutional and regulatory details of the U.S. equities market, shows that under the status
quo market design: (i) trading behavior across the many distinct exchanges is as if there is just a single
“synthesized” exchange, as opposed to traditional platform competition; (ii) as a result, trading fees are
perfectly competitive; but (iii) exchanges capture and maintain significant economic rents from the sale
of “speed technology” (i.e., proprietary data feeds and co-location)—arms for the arms race. Using a
variety of data, we document seven stylized empirical facts that suggest that the model captures the
essential economics of how U.S. stock exchanges compete and make money in the modern era. We then
use the model to examine the private and social incentives for market design innovation. We show that
the market design adoption game among incumbent exchanges is not a coordination game, but rather
a repeated prisoner’s dilemma. If an exchange adopts a new market design that eliminates latency
arbitrage, it would win share and earn economic rents. However, imitation by other exchanges would
result in an equilibrium that resembles the status quo with competitive trading fees, but now without
the rents from the speed race. This means that although the social returns to market design innovation
are large, the private returns are much smaller and may be negative, especially for incumbents that
derive rents from the status quo. Despite this negative result, however, our analysis does not imply that
a market-wide market design mandate is necessary to fix the problem. Rather, it suggests a modest
regulatory “push” may be sufficient to tip the balance of incentives and encourage “the market to fix the
market.”
∗ Project start date: April 2015. We are especially grateful to Larry Glosten and Terry Hendershott for serving as discussants
of an early version of this project. We also thank Jason Abaluck, Nikhil Agarwal, Susan Athey, John Campbell, Dennis
Carlton, Judy Chevalier, John Cochrane, Christopher Conlon, Peter Cramton, Doug Diamond, David Easley, Alex Frankel,
Joel Hasbrouck, Kate Ho, Anil Kashyap, Pete Kyle, Donald Mackenzie, Neale Mahoney, Paul Milgrom, Joshua Mollner, Ariel
Pakes, Al Roth, Fiona Scott Morton, Andrei Shleifer, Jeremy Stein, Mike Whinston, Heidi Williams, Luigi Zingales, and
numerous industry practitioners and seminar participants for helpful discussions and suggestions. Paul Kim, Cameron Taylor,
Matthew O’Keefe, Natalia Drozdoff, and Ethan Che provided exceptional research assistance. Budish acknowledges financial
support from the Fama-Miller Center, the Stigler Center, and the University of Chicago Booth School of Business. Disclosure:
the authors declare that they have no relevant or material financial interests that relate to the research described in this paper.
John Shim worked at Jump Trading, a high-frequency trading firm, from 2006-2011.
† University of Chicago Booth School of Business and NBER, eric.budish@chicagobooth.edu
‡ Harvard University and NBER, robinlee@fas.harvard.edu
§ University of Chicago Booth School of Business, john.shim@chicagobooth.edu
1 Introduction
“We must consider, for example, whether the increasingly expensive search for speed has passed
the point of diminishing returns. I am personally wary of prescriptive regulation that attempts
to identify an optimal trading speed, but I am receptive to more flexible, competitive solutions
that could be adopted by trading venues. These could include frequent batch auctions or other
mechanisms designed to minimize speed advantages. . . . A key question is whether trading venues
have sufficient opportunity and flexibility to innovate successfully with initiatives that seek to
deemphasize speed as a key to trading success in order to further serve the interests of investors. If
not, we must reconsider the SEC rules and market practices that stand in the way.” (Securities
and Exchange Commission Chair Mary Jo White, June 2014)
As of early 2019 there are 13 stock exchanges in the U.S., across which over 1 trillion shares ($50 trillion)
are traded annually. All 13 exchanges use a market design called the continuous limit order book. A recent
paper of Budish, Cramton and Shim (2015) showed that this market design has an important design flaw.
The combination of (i) treating time as a continuous variable, and (ii) processing requests to trade serially,
causes latency arbitrage—defined as arbitrage rents from symmetrically observed public information—to
be a built-in equilibrium feature of the market design. Latency arbitrage causes markets to be less liquid
than they could be, leads to a never-ending and socially-wasteful arms race for speed, and offends common
economic intuitions about what constitutes an efficient market. Budish, Cramton and Shim showed that the
fix is conceptually pretty simple, requiring just two modifications to the continuous limit order book: (i)
treat time as a discrete variable (analogously to prices, which come in discrete units); and (ii) in the event
that multiple orders arrive at the same discrete time, batch process them using a standard uniform-price
auction. However, despite the remarks of the SEC Chair above and encouragement from others,1 the main
U.S. stock exchanges have not shown much interest in changing their market design. In the two instances
where a startup (IEX) and a small exchange (CHX) proposed market design ideas similarly motivated by
concerns about aspects of high-frequency trading, the proposals were met with fierce resistance from the
large incumbent exchanges and from many high-frequency trading firms.2,3
This begs the question: are private forces alone sufficient to foster the adoption of innovative and more
1 See New York Attorney General Schneiderman (2014); Bloomberg Editorial Board (2014), which expressed views of both
Bloomberg’s editors and Goldman Sachs; and current Federal Reserve Chair (then Governor) Powell (2015), who in the context
of the U.S. Treasury market remarked “Ideas such as these make me wonder whether it might collectively be possible to come
to a compromise in which more trading is done directly on the public market, if at the same time the public market rules were
adjusted to emphasize greater liquidity provision, and particularly more stable liquidity provision, over speed.”
2 Our sense is that the resistance to the Investors’ Exchange (IEX)—an unprecedented 477 SEC comment letters were filed
regarding its exchange application, including one in which the New York Stock Exchange wrote “Like the ‘non-fat yogurt’
shop on Seinfeld, which actually serves tastier, full-fat yogurt to increase its sales, IEX advertises that it is ‘A Fair, Simple,
Transparent Market,’ whereas it proposes rules that would make IEX an unfair, complex, and opaque exchange” (NYSE,
2015b)—reflected a genuine mixture of incumbents’ desire to preserve the status quo and legitimate concerns about the details
of IEX’s market design. The most important concern, from the perspective of the current paper, is that IEX’s market design only
protects against latency arbitrage for non-displayed pegged orders; it does not protect against latency arbitrage for conventional
displayed limit orders. Its displayed (“lit”) market is identical to a standard continuous limit order book, just 350 microseconds
further away from market participants than would be the case from geography alone, due to the famous “speed bump”. Please
see Budish (2016b) for further details.
3 The Chicago Stock Exchange (CHX) proposed to adopt an “asymmetric” speed bump for its exchange—in which liquidity
taking orders are delayed but liquidity providing orders are not delayed—which would have protected against latency arbitrage in
the displayed market, unlike IEX’s symmetric speed bump. But it, too, met with fierce resistance from the larger exchanges and
several high-frequency trading firms—for example, Citadel wrote that it “unfairly structurally and systematically discriminates
against market participants that are primarily liquidity takers.” (Citadel, 2016) CHX ultimately withdrew the proposal and
was acquired by the New York Stock Exchange (Michaels and Osipovich, 2018). Please see Baldauf and Mollner (2018a) for a
detailed theoretical analysis of asymmetric speed bumps, Section VIII.C-D of Budish, Cramton and Shim (2015) for additional
discussion of speed bumps, and Budish (2016a) for further details of CHX’s proposal.
1
efficient market designs? Implicit in the quote at the top of the paper—delivered in a speech by then SEC
Chair Mary Jo White—is the view that private and social incentives for market design innovation are aligned:
if there is a market design innovation that is efficiency enhancing, then private market forces will naturally
evolve towards realizing the efficiency if allowed to do so (Griliches, 1957). However, as is well known, there
are numerous economic settings where private and social incentives for innovation diverge (Arrow, 1962;
Nordhaus, 1969; Hirshleifer, 1971). The question of whether or not sufficient innovation incentives exist
for stock exchanges is timely and of significant importance. If they do, then it follows that “prescriptive
regulation” should not mandate a specific market design; rather, regulators should ensure that regulation
does not “stand in the way” of “competitive solutions.” If not, intervention may be warranted—all the more
so if the potential economic savings are substantial.4
In this paper, we argue that incumbent exchanges do not have sufficient incentives to adopt new market
designs precisely because they derive rents from the inefficiencies that these alternative designs seek to elim-
inate. That is, even though there are multiple exchanges that appear to compete fiercely with one another
for trading volume, they—alongside high-frequency trading firms and speed-technology providers—capture
and maintain a significant share of the economic rents at stake in the speed race. We emphasize that our
story is not one of liquidity externalities, multiple equilibria due to coordination failure, chicken-and-egg,
etc., as is central in the literature on network effects and platform competition (e.g., Farrell and Saloner
(1985); Katz and Shapiro (1986); Rochet and Tirole (2003); Farrell and Klemperer (2007)) and past market
microstructure literature on financial exchange competition (cf. surveys by Madhavan (2000) and Cantillon
and Yin (2011)). Rather, our story in the end is ultimately a more traditional economic one of incumbents
protecting rents and missing incentives for innovation.
Central to our argument is a novel theoretical model of the stock exchange industry, tailored to the
institutional and regulatory details that shape modern electronic trading, and built to understand the nature
of exchange competition and the associated incentives for innovation. There are four types of players in our
model, all strategic: exchanges, trading firms, investors, and informed traders. Initially, mirroring the status
quo of the current market, we assume that all exchanges employ the continuous limit order book market
design. Exchanges are undifferentiated, and strategically set two prices: per-share trading fees, and fees
for “speed technology” that enables trading firms to receive information about and respond more quickly
to trading opportunities on a given exchange. In practice, speed technology includes co-location (the right
to locate one’s own servers right next to the exchange’s servers) and proprietary data feeds (which enable
trading firms to receive updates from the exchange faster than from non-proprietary data feeds). Trading
firms choose the set of exchanges to buy speed technology from. They also choose whether and how to provide
liquidity by choosing the exchange(s) on which to offer liquidity, the quantity to offer on each exchange, and
a bid-ask spread on each exchange. The bid-ask spread trades off the benefits of providing liquidity to
investors (thereby collecting the spread) versus the cost of either being adversely selected against by an
informed trader (as in Glosten and Milgrom (1985)) or being on the losing end of a latency arbitrage race
with other trading firms—i.e., being “sniped” (as in Budish, Cramton and Shim (2015)).
Our analysis of the status quo delivers three main results. First, as in Glosten (1994), although the
market can be fragmented in the sense that trading activity is split across several exchanges, economically
many aspects of trading activity behave as if there is just a single “synthesized” exchange.5 Specifically: all
4 While there is no exact consensus in the academic literature on the economic stakes in the high-frequency trading arms
race, the academic estimates that are available suggest that it is in the single-digit billions of dollars per year for U.S. equities.
Taking a net present value of this amount, and extrapolating across countries and other financial instruments, it is easy to get
to a net present value figure in excess of $100 billion. See Budish (2017) for discussion.
5 Glosten (1994) presciently foresaw that frictionless search and order-splitting across electronic markets (cf. his Assumption
2
liquidity is at the same prices and bid-ask spreads regardless of the exchange on which it is offered, with the
marginal unit of liquidity indifferent across exchanges due to a linear relationship between the quantity of
liquidity on an exchange (i.e., market depth) and the quantity of trade on that exchange (i.e., volume); and
aggregate depth and volume are both invariant to how trading activity is allocated across exchanges. This
behavior is brought about by two key sets of regulations in the U.S.: Unlisted Trading Privileges (UTP) and
Regulation National Market System (Reg NMS).6 UTP essentially implies that stocks are perfectly fungible
across exchanges: i.e., a stock that is technically listed on exchange X can be bought on any exchange Y and
then sold on any exchange Z. Reg NMS ensures that searching among exchanges, and then transacting across
(“accessing”) them, are both frictionless. This frictionless search and access allows market participants to
costlessly “stitch together” the order books across the various exchanges, and yields investor demand that
is perfectly responsive to price differences across exchanges. This behavior also leads to our second result:
due to the same frictionless search and access, investor demand is perfectly elastic with respect to trading
fees as well; hence, fierce Bertrand-style competition yields competitive (zero) trading fees on all exchanges.
As intuition for the first two results, consider a hypothetical world with buyers and sellers of a single
good, and multiple platforms on which transactions can occur. A regulation corresponding to UTP would
ensure that this good is perfectly homogeneous—e.g., no small differences between the types of drivers on
Uber versus Lyft—and can be bought or sold on any platform. A regulation corresponding to Reg NMS
ensures that searching for the best price across platforms and then potentially engaging in a transaction are
literally frictionless—e.g., not 10 extra seconds to check a second ride-sharing app or 10 minutes to drive to
a store, but no time at all. Given this, it is intuitive to see why: (i) aggregate economic activity will not
depend on how sellers allocate their goods across platforms (as buyers will find sellers, regardless of where
they are); and (ii) platform transaction fees will be Bertrand-competed down to the competitive level. There
is a fundamental economic difference between an “almost” commodity and “cheap” search, and an identical
commodity and zero-cost search (cf. Diamond (1971)).
Our third result is that exchanges can both capture and maintain substantial rents from the sale of speed
technology. This may appear surprising as exchanges are modeled as undifferentiated and search and access
is frictionless; as we have mentioned, these same features lead to competitive trading fees. There are two
reasons why exchanges earn supra-competitive rents for speed technology in equilibrium. First, even though
stocks are fungible across exchanges, latency-sensitive trading opportunities are not: if there is a sniping
opportunity that involves a stale quote on Exchange X, only trading firms that have purchased Exchange X
speed technology will be able to effectively compete in the sniping race. As long as trading firms multi-home
and purchase speed technology from all exchanges (which they do in equilibrium), exchanges can charge
positive fees for speed technology without incentives to undercut each other. Second, in contrast to basic
models of add-on pricing whereby profits from add-on goods are dissipated by firms selling the primary
good below cost (cf. Ellison (2005); Gabaix and Laibson (2006)), exchange rents earned from the sale of
speed technology are not dissipated via further competition on trading fees. The reason is that trading fees
are already at zero, and cannot become negative without creating a “money-pump” wherein trading firms
execute infinite volume to extract the negative fee.
We also prove that although exchanges are modeled as price setters who post take-it-or-leave-it offers
to trading firms for speed technology, exchanges nevertheless cannot extract all of the industry rents from
4) could generate what we refer to as the single synthesized exchange (cf. his Proposition 8), well over a decade before the
passage of Reg NMS. Please see Section 3.2.4 for a detailed discussion of the relationship between Glosten (1994) and this
aspect of our analysis.
6 These regulations are described in detail in Section 2.
3
latency arbitrage.7 The reason is that trading firms are able to influence where volume is transacted, and
this allows them to discipline exchanges that attempt to take too much of the pie.8
Although our model is highly stylized and abstracts from several real-world-complications (which include
agency frictions, tick-sizes, asymmetric trading fees, and strategic trading over time as in Kyle-style models),
we establish that our parsimonious model nonetheless does reasonably well in matching several empirical
moments found in the data. Specifically, using a combination of trades-and-quotes (TAQ) data and exchange-
company financial filings (e.g., 10-K’s, S-1’s, merger proxies, fee filings), we document seven stylized facts
about modern-era stock exchange competition that align with the model. Our first series of facts relates to
our result that the market behaves as if trading activity occurred on a single synthesized exchange. These
facts, documented using a sample of reasonably highly traded stocks, include all major exchanges typically
having displayed liquidity at the same best price; a close linear relationship between the quantity of liquidity
on an exchange (i.e., its displayed depth) and its trading volume; and market shares that are interior (i.e., no
tipping). Next, we document that trading fees across major exchanges are economically very small. Trading
fees are quite complicated (cf. Chao, Yao and Ye (2019)), but using a variety of data sources to cut through
this complexity, we compute that the average fee for regular-hours trading, across the three largest stock
exchange families, is around $0.0001 per share per side—or about 0.0001% per side for a $100 stock. This
implies that across approximately 1 trillion shares traded during regular hours each year, exchanges earn
approximately $200 million in trading fees. To put this in perspective, StubHub, the largest secondary-
market venue for concert and sports tickets, has revenues exceeding $1 billion; that is, StubHub’s revenue is
over five times that for all U.S. regular-hours equities trading, despite the secondary market for event tickets
being a tiny fraction of the secondary market for U.S. equities. Last, we document that exchanges earn
significant revenues from the sale of co-location services and proprietary data feeds. For the BATS exchange
family, for which the data is the cleanest, revenue from co-location and data is about 69% of total revenue.
In aggregate across the three major exchange families (BATS, Nasdaq, NYSE), we document significant
growth in ESST fees during the Reg NMS era (post 2007), with 2018 speed technology revenues estimated
to be on the order of $1 billion.
Overall, the empirical facts suggest that our simple model—though explicitly abstracting away from
many aspects of modern U.S. stock trading—captures the essential economics of how U.S. stock exchanges
compete and make money in the modern era. Also importantly, the empirical facts we document for the
U.S. stock exchange industry, taken in total, are not consistent with many other models of financial ex-
change competition in prior literature. These include models that feature single-homing, network effects,
market tipping, supra-competitive trading fees, and so forth (e.g., Pagano (1989); Ellison and Fudenberg
(2003); Cantillon and Yin (2008)); models in which exchanges are meaningfully horizontally or vertically
differentiated (e.g., Baldauf and Mollner (2018b); Pagnotta and Philippon (2018)); and models in which
tick-size frictions are central (Chao, Yao and Ye (2017, 2019)). Nor are these facts consistent with standard
models of platform competition from other economic contexts (cf. Rochet and Tirole (2003, 2006); Farrell
and Klemperer (2007)).
7 Taking our bound literally, and using realistic parameters for the numbers of fast trading firms and exchanges, our model
suggests that exchanges in aggregate can extract at most about 20% of the total latency arbitrage prize. Please see Section
3.2.4 for discussion.
8 A particularly extreme version of this move was announced very recently in January 2019 as several large high-frequency
trading firms and broker-dealers announced that they were exploring starting a new exchange, called MEMX, out of concern
about rising co-location and proprietary data fees (Osipovich, 2019b). The financial columnist Matt Levine wrote: “While
the last new stock exchange to launch in the U.S., the Investors Exchange or IEX, was self-consciously about protecting long-
term fundamental investors from the ravages of high-frequency trading, MEMX seems to be self-consciously about protecting
high-frequency traders from the ravages of stock-exchange fees” (Levine, 2019).
4The last part of our analysis uses the model to address our overarching question: will the market adopt
new market designs, such as frequent batch auctions (Budish, Cramton and Shim, 2015), that address the
negative aspects of high-frequency trading? How do exchanges’ private innovation incentives relate to social
incentives? Our model suggests that exchanges are unlikely to embrace such innovation with open arms.
However, it is not because a new market design that eliminates latency arbitrage would fail to be utilized
by market participants—indeed, if introduced, it would gain significant market share—but rather because
such innovation would destroy the rents that incumbent exchanges currently earn from speed technology. To
conduct this analysis, we extend our theoretical model to allow for exchanges to operate one of two market
designs: either the continuous-time limit order book (Continuous), or discrete-time frequent batch auctions
(Discrete). Importantly, in the context of competition with the Continuous market, we consider frequent
batch auctions with a very short batch interval: long enough to effectively batch process if multiple trading
firms react to the same public signal at the same time, but otherwise essentially as short as possible.9
We show that if only one exchange employs Discrete while all others employ Continuous, the Discrete
exchange will capture a large share of trading volume and large economic rents. Intuitively, eliminating
latency arbitrage eliminates a tax on liquidity, and the fact that market participants can frictionlessly access
and search across exchanges ensures that if there are two markets operating in parallel, one with a tax and
one without, the one without the tax will take off. That is, the standard coordination problems associated
with getting a new market off the ground do not apply here, and the Discrete exchange is able to earn trading
fees commensurate with the tax that it eliminates.10,11 However, this frictionless world is also a double-edged
sword: although it rewards a first-mover, it also implies that any subsequent adoption of Discrete by other
exchanges—which our model suggests is likely—leads to the same Bertrand competition on trading fees as
before, but now without the industry rents from the speed race. Thus, we establish that the market design
adoption game among incumbent exchanges is essentially a repeated prisoner’s dilemma: while any one
exchange has incentive to unilaterally “deviate” and adopt Discrete, all incumbents prefer the Continuous
status quo, in which they share in latency arbitrage rents, to a world in which all exchanges are Discrete, and
these rents are gone. Similar arguments imply that a de novo entrant exchange using Discrete would have
difficulty recouping any substantial fixed costs of entry if imitation by other exchanges is likely and rapid.
Thus, we conclude that private incentives to adopt a new market design that eliminates latency arbitrage
are dramatically lower than social incentives, and potentially even negative.
Finally, we discuss policy implications. The basic question is whether (i) there will be a private-market
solution to latency arbitrage and the arms race (i.e., “will the market fix the market”), or (ii) would some
sort of regulatory intervention—ranging from a market-wide market design mandate to something more
circumscribed—be required. Our analysis suggests that although private incentives alone may be insufficient,
9 In practice, given advances in speed technology over the last several years, 1 millisecond would likely be more than sufficient
to effectively batch process; some industry participants have argued to us that as little as 50 microseconds (i.e., 0.000050 seconds)
might suffice. A batch interval of 1 millisecond or less would also allow the frequent batch auction exchange to operate within
the framework of Reg NMS, which is significant. See Section 2.2 for additional discussion of Reg NMS. See Section 5 for the full
details of how we model frequent batch auctions, including the important details regarding information policy which, following
Budish, Cramton and Shim (2015), is analogous to information policy in the continuous market but with the same information
(about trades, cancels, the state of the order book, etc.) disseminated in discrete time, at the end of each interval.
10 This result may seem to contradict the result in Glosten (1994), Proposition 9, that finds that the electronic limit order
book is in a certain sense “competition proof.” The explanation is that the Glosten (1994) model implicitly precludes the
possibility of latency arbitrage. The reason Discrete “wins” against Continuous in our model is precisely because it eliminates
latency arbitrage. Please see Section 5.1.2 for further discussion.
11 In our model, with continuous prices, the unique equilibrium is for Discrete to win 100% market share as long as the sniping
tax is strictly positive. In a richer model, with tick-size constraints (i.e., a restriction that prices must be in increments of $0.01),
if the sniping cost per share is smaller than the tick size—which seems empirically to be the case—then there also may exist
equilibria without complete tipping in which Discrete’s market share depends on the ratio of sniping costs per share to the tick
size. Please see Section 5.1.2 for further discussion.
5a modest regulatory “push”—one that tips the balance of incentives enough to get a de novo exchange to
enter or an incumbent to adopt—might suffice. The rough intuition for why a push may suffice, as opposed
to needing an all-out mandate, is that investors strictly prefer markets without the latency arbitrage tax, and
investors are ultimately who exchanges and trading firms make their money from—so, once such a market
enters, private-market forces can take over. Such pushes might include: (i) reducing the entry and adoption
costs of launching a new stock exchange, for example, by lowering the risk of failed entry by clarifying which
types of market designs would be admissible under Reg NMS, or finding some way to subsidize the fixed
costs of entry; and (ii) a modest regulatory exclusivity period for the innovator, during which competing
exchanges would not be able to imitate the design. Analogous to FDA exclusivity periods for non-patentable
drugs, an SEC exclusivity period could induce a first-mover exchange to invest the fixed costs associated
with developing, implementing and gaining regulatory approval for a new market design.
Our paper makes several contributions to the literature. First is our theoretical industrial organization
model of the stock exchange industry, described by some as the single most iconic market in global cap-
italism (e.g., 60 Minutes (2014)). We depart from much of the previous literature on financial exchange
competition in both our focus—the source of economic profits for U.S. stock exchanges and their incentives
to adopt innovative market designs—and in our modeling approach. Most centrally, most other papers in
this literature have some sort of single-homing, either by market participants choosing which one exchange
to trade on (e.g., Pagano (1989); Santos and Scheinkman (2001); Ellison and Fudenberg (2003); Pagnotta
and Philippon (2018); Baldauf and Mollner (2018b)), or by financial instruments that are specific to a single
exchange (as in Cantillon and Yin (2008)). This single-homing is often (though not always) accompanied
by some meaningful differentiation across exchanges, either horizontally or vertically. By contrast, in our
model, motivated by the regulatory environment for modern electronic U.S. stock trading, stocks are fungi-
ble across exchanges, market participants can frictionlessly multi-home across exchanges, and exchanges are
undifferentiated. This modeling approach also leads to “economics of the status quo” that are fundamen-
tally different from those that would emerge under standard platform or two-sided competition frameworks
where, typically, platforms earn rents from platform-specific network effects by charging supra-competitive
access and transaction fees (cf. Caillaud and Jullien (2003); Rochet and Tirole (2003); Armstrong (2006);
Farrell and Klemperer (2007)). Here, since exchanges are modeled as undifferentiated and exchange-specific
network effects are nullified due to frictionless search and access, trading fees are competitive—zero in our
model, and approximately zero in the data. A related insight of our model that may be of interest to the
platforms literature is that, while the market may appear to be fragmented across multiple exchanges, the
market behaves in some respects as if there were a single “synthesized” exchange. The market microstruc-
ture literature has in the past been puzzled by fragmentation (cf. Madhavan (2000) and what he terms
the “Network Externality Puzzle”). Here, we provide a theoretical rationale for why fragmentation per se
may not necessarily lead to trading inefficiencies—this aspect of our analysis builds on a prescient result of
Glosten (1994) and aligns with empirical evidence in O’Hara and Ye (2011).
There are also two technical features of our theoretical analysis worth highlighting. First, in our analysis
we develop and motivate an equilibrium solution concept which we refer to as an order-book equilibrium
to address equilibrium existence issues. This solution concept is closely related to alternative solution
concepts employed in the insurance market literature (e.g., Wilson (1977); Riley (1979)), and may prove
useful analyzing other markets with adverse selection. Second, we generate a strictly interior split of latency
arbitrage rents between exchanges and trading firms without relying on an explicit bargaining model; we
show that this arises as a result of exchanges being able to post prices for speed technology (which they do in
6reality), and trading firms being able to steer trading volume via the provision of liquidity (which they can
in reality). Though we acknowledge that these particular contributions (and our modeling exercise overall)
may be highly tailored for a specific market, we believe that this specificity is justified by the importance of
the industry.
Our paper’s second contribution is the seven stylized facts that, to our knowledge, have not been doc-
umented in this form elsewhere. In particular, the facts on trading fees and on speed technology fees may
be of direct use for current policy debates. The SEC recently announced a pilot study on transaction fees,
focusing on the controversial practice of “maker-taker” fee-and-rebate pricing models (U.S. Securities and
Exchange Commission, 2018c). While our results do not speak to the agency concerns at the heart of the
controversy (cf. Battalio, Corwin and Jennings (2016)), our results do show that, once one cuts through the
complexity of modern fee schedules, the average fees are economically small. With respect to speed technol-
ogy fees, in October 2018 for the first time in recent history the SEC rejected proposed data fee increases
by NYSE and Nasdaq (Clayton, 2018). In a speech around that time Commissioner Robert J. Jackson Jr.
called for “greater transparency about how exchanges make their money. . . and a clear and uniform approach
to disclosing revenues across exchanges and over time.” He described that he and his staff “tried and failed
to use public disclosures to meaningfully examine exchanges’ businesses. . . [and] attempted to look into the
revenues that exchanges generate from selling market data and connectivity services. We expected that
such numbers would be available. . . but found. . . it nearly impossible” (Jackson Jr., 2018). Our estimate of
total exchange speed-technology revenues—which, as the reader will see, triangulates from numerous data
sources in lieu of obvious, transparent numbers from exchange filings—is surely not perfect, but it provides
a magnitude that market policy makers currently lack.
Last is our analysis of the question “will the market fix the market?” More precisely, the intellectual
contribution is in using the model to fill in the cells of the market design adoption game payoff matrix; once
we understand that the adoption game constitutes a prisoner’s dilemma (as opposed to, e.g., a coordination
game), the rest of the analysis and discussion is straightforward. We use these insights to identity a modest
policy response, well short of the “prescriptive regulation” that SEC Chair White expressed wariness of.
We view this particular contribution as in the spirit of economic engineering (Roth, 2002), working with
the real-world constraints of the specific market design setting, rather than assuming the ability to design
institutions from scratch.12
Roadmap. The remainder of this paper is organized as follows. In Section 2, we describe the key institu-
tional features of the stock market that shape our theoretical model of exchange competition. We introduce
and analyze the model of the status quo in Section 3. Section 4 provides our seven stylized empirical facts. In
Section 5, we use our theoretical model to examine competition among alternative market designs. Section
6 proposes potential policy responses, and Section 7 concludes.
2 Institutional Background
Readers of this paper—especially researchers who are less familiar with financial market microstructure—
may have in mind, when thinking of stock exchanges and how they compete, the old New York Stock
12 In this spirit, our paper belongs to an active literature at the intersection of finance and market design, with some recent
works including Antill and Duffie (2018), Brogaard, Hendershott and Riordan (2017), Bulow and Klemperer (2013, 2015),
Du and Zhu (2017), Duffie and Dworczak (2018), Duffie and Zhu (2016), Glode and Opp (2016), Hendershott and Madhavan
(2015), Hortaçsu, Kastl and Zhang (2018), Kastl (2017), Kyle and Lee (2017), Kyle, Obizhaeva and Wang (2018), and Menkveld,
Yueshen and Zhu (2017).
7Exchange floor. As recently as the 1990s, if a stock was listed on the New York Stock Exchange, the large
majority of its trading volume (65% in 1992) transacted on the New York Stock Exchange floor. Similarly,
if a stock was listed on Nasdaq,13 a large majority of its volume transacted on the Nasdaq exchange (86%
in 1993).14 In this earlier era, stock exchanges enjoyed valuable network effects and supra-competitive fees.
The seminal model of Pagano (1989)—in which traders single-home, and there are liquidity externalities that
can cause traders to agglomerate on an exchange with supra-competitive fees—was a reasonable benchmark
for thinking about the industrial organization of the industry (see also Ellison and Fudenberg, 2003).
This model, however, is less applicable for the modern era of stock trading.15 In our data, from 2015,
there are 12 exchanges, all stocks trade essentially everywhere, and market shares are both stable and interior
(i.e., no tipping). There are 5 exchanges with greater than 10% market share each (83% in total), and the
next 3 exchanges together have another 15% share. Please see our discussion of Stylized Fact #3 in Section
4.1 for further details. Trading fees, while quite complex and in many ways opaque (cf. Chao, Yao and Ye
(2019)), are ultimately quite small, as we will document rigorously as Stylized Fact #4 in Section 4.2.
There are two key sets of regulations that together shape the industrial organization of modern elec-
tronic stock exchanges. The first set, related to Unlisted Trading Privileges (UTP), has its roots in the 1934
Exchange Act and in its modern incarnation enables all stocks to trade on all exchanges, essentially inde-
pendently of where the stock is technically listed. The second set, Regulation National Market System (Reg
NMS), was implemented in 2007 and requires that information about trading opportunities (i.e., quotes)
be automatically disseminated across the whole market (including both other exchanges and entities such
as brokers), and also requires, roughly, that the whole market pay attention to such information and direct
trades to the most attractive prices across the whole system. As we will see in our formal model in Section
3, this effectively nullifies any exchange-specific network effects.16
In this institutional background section we describe each of these sets of regulations; our goal is to provide
a level of detail that is sufficient to justify our modeling choices.
We note that while our discussion focuses on the United States, there are economically similar regulations
for stock exchanges in Canada and somewhat similar regulations in Europe.17 Regulations for futures
exchanges, on the other hand, are quite different from those for stock exchanges, both in the U.S. and
abroad. In particular, there is no analogue of UTP in futures markets because each contract is proprietary
to a particular exchange. This has a significant effect on the industrial organization of futures markets as
distinct from stock markets, as we will discuss briefly under Stylized Fact #6 in Section 4.3 (see especially
Figure 4.5). Similarly, there are differences between the regulation of stock exchanges and the regulation of
financial exchanges for other financial instruments like government bonds, corporate bonds, foreign currency,
13 Technically, stocks could not be “listed” on Nasdaq until it became an exchange in 2006, but the 1975 Exchange Act
Amendments enabled stocks to trade over-the-counter via Nasdaq achieving something economically similar.
14 For the NYSE market share claim, see the SEC study “Market 2000”, Exhibit 18 (U.S. Securities and Exchange Commission,
1994). For the Nasdaq market share claim, see the SEC Market 2000 study, Exhibit 12.
15 For surveys of modern electronic trading, focusing on a broader set of issues than stock exchanges per se, good starting
points are Jones (2013), Fox, Glosten and Rauterberg (2015, 2019), O’Hara (2015) and Menkveld (2016).
16 Note that “dark pools”, or Alternative Trading Systems, are not governed by Reg NMS. Instead, dark pools typically
facilitate trade at prices that reference the best available quotes from exchanges (e.g., at the midpoint). This of course raises
its own interesting economic issues, specifically that dark pools may “free ride” off of prices discovered by the exchanges. See,
for instance, Hendershott and Mendelson (2000), Zhu (2014), and Antill and Duffie (2018). A good topic for future research
would be to incorporate latency arbitrage into a model with competition between exchanges and dark pools.
17 In Canada’s version of the Order Protection Rule (which goes by the same name), the key difference is that the rule applies
to the full depth of the order book, not just the first level (Canadian Securities Administrators, 2009). In Europe, instead of
the (prescriptive) Order Protection Rule there are (principles-based) best execution regulations (Petrella, 2010). Note however
that principles-based best execution requirements leave some ambiguity with regard to whether market participants have to
“pay attention” to quotes from small exchanges, which could affect innovation incentives; whereas under the Order Protection
Rule there is no such ambiguity. This seems a good topic for future research.
8etc.; in particular, the information dissemination provisions of Reg NMS are often economically different in
these asset classes. Again, our focus will be on U.S. stock exchanges, though we think there is much interesting
future research to do on the industrial organization of financial exchanges for other kinds of assets, under
other regulatory regimes, and so forth.
2.1 Unlisted Trading Privileges (UTP)
Section 12(f) of the 1934 Exchange Act (15 U.S.C. 78a, 1934), passed by Congress, directed the Securities
and Exchange Commission to “make a study of trading in unlisted securities upon exchanges and to report
the results of its study and its recommendations to Congress.” Since that time, the right of one exchange
to facilitate trading in securities that are listed on other exchanges has undergone several evolutions. In its
current form, passed by Congress in the Unlisted Trading Privileges Act of 1994 (H.R. 4535, U.S. Congress,
1994) and clarified by the SEC in a Final Rule effective November 2000 (U.S. Securities and Exchange
Commission, 2000), one exchange may extend unlisted trading privileges (UTP) to a security listed on
another exchange immediately upon the security’s initial public offering on the listing exchange, without any
formal application or approval process through the SEC. Prior to 1994, exchanges had to formally apply to the
SEC for the right to extend UTP to a particular security; such approval was “virtually automatic” following
a delay of about 30-45 days (Hasbrouck, Sofianos and Sosebee, 1993). Between the passage of the UTP Act
of 1994 and the Final Rule in 2000, extension of UTP was automatic but only after an initially two-day, and
then one-day, delay period after the security first began trading on its listing exchange (U.S. Securities and
Exchange Commission, 2000). For further historical discussion of UTP, please see the background section of
the 2000 Final Rule document, and also Amihud and Mendelson (1996).
For the purposes of our theoretical model, we will incorporate UTP in its current form by assuming that
the security in the model is perfectly fungible across exchanges. This captures that regardless of where a
security is listed, was last traded, etc., it can be bought or sold on any exchange, and its value is the same
regardless of where it is traded.
2.2 Regulation National Market System (Reg NMS)
Regulation National Market System (“Reg NMS”, U.S. Securities and Exchange Commission, 2005) passed
in June 2005 and implemented beginning in October 2007, is a long and complex piece of regulation, with
routes tracing to the Securities Exchange Act Amendments of 1975 and the SEC’s “Order Handling Rules”
promulgated in 1996.18 For the purpose of the present paper, however, there are two core features to
highlight.19
The first is the Order Protection Rule, or Rule 611. The Order Protection Rule prohibits an exchange
from executing a trade at a price that is inferior to that of a “protected quote” on another exchange.
A quote on a particular exchange is “protected” if it is (i) at that exchange’s current best bid or offer;
and (ii) “immediately and automatically accessible” by other exchanges. Reg NMS does not provide a
18 The goal of the National Market System is described by the SEC as follows: “The NMS is premised on promoting fair
competition among individual markets, while at the same time assuring that all of these markets are linked together, through
facilities and rules, in a unified system that promotes interaction among the orders of buyers and sellers in a particular
NMS stock. The NMS thereby incorporates two distinct types of competition—competition among individual markets and
competition among individual orders—that together contribute to efficient markets.” U.S. Securities and Exchange Commission
(2005, pg 12)
19 For an overview of Reg NMS, a good source is the introductory section of the SEC’s final ruling itself (U.S. Securities and
Exchange Commission, 2005). For an overview of the National Market System prior to Reg NMS, good sources are O’Hara and
Macey (1997) and the SEC’s “Market 2000” study (U.S. Securities and Exchange Commission, 1994).
9precise definition of “immediately and automatically accessible,” but the phrase certainly included automated
electronic continuous limit order book markets and certainly excluded the NYSE floor system with human
brokers.20
A June 2016 rules clarification issued by the SEC indicated that exchanges can use market designs that
impose delays on the processing of orders and still qualify as “immediate and automatic” so long as (i) the
delay is of a de minimis level of 1 millisecond or less, and (ii) the purpose of the delay is consistent with the
efficiency and fairness goals of the 1934 Exchange Act (U.S. Securities and Exchange Commission, 2016b).
This rules clarification suggests that quotes in a frequent batch auction exchange would be protected under
Rule 611 so long as the batch interval was no longer than the de minimis threshold of 1 millisecond; however,
this specific market design has not yet been put before the SEC for explicit approval.21
An additional detail about the Order Protection Rule that bears emphasis is that, in practice, sophisti-
cated market participants can take on responsibility for compliance with the Order Protection Rule them-
selves, absolving exchanges of the responsibility. They do so using what are known as intermarket sweep
orders, or ISOs. If an exchange receives an order that is not marked as ISO, then it is the exchange’s re-
sponsibility to ensure that it handles the order in a manner compliant with the Order Protection Rule (e.g.,
it cannot execute a trade that trades through a protected quote elsewhere). If an exchange receives an order
that is marked as ISO, then the exchange may presume that the sender of the order has ensured compliance
with the Order Protection Rule (e.g., by also sending orders to other exchanges to attempt to trade with any
relevant protected quotes) and the exchange need not check quotes elsewhere before processing the order.22
The second key provision to highlight is the Access Rule, or Rule 610. Intuitively, for an exchange to be
able to comply with the Order Protection Rule it must be able to efficiently obtain the necessary information
about quotes on other exchanges, and, if necessary, be able to efficiently route orders to trade against quotes
on other exchanges. Similarly, a broker-dealer seeking to comply with the Order Protection Rule using ISOs
must be able to efficiently obtain information about quotes from all exchanges, and efficiently trade against
quotes on all exchanges. As the SEC writes (pg. 26), “. . . protecting the best displayed prices against trade-
throughs would be futile if broker-dealers and trading centers were unable to access those prices fairly and
efficiently.”
The Access Rule has three sets of provisions that together are aimed at ensuring such efficient access—or
what we will sometimes call “search and access,” to highlight that economically the Access Rule (and related
rules that affect information provision, such as those governing slower, non-proprietary market data feeds)23
20 A central issue in the debate over IEX’s exchange application was whether IEX’s quotes, given that its market design
included a “speed bump”, would count as immediately and automatically accessible under Reg NMS. See U.S. Securities and
Exchange Commission (2016a) for the (unprecedented number of) public comments on IEX’s exchange application. See also
footnote 2 for additional details regarding IEX’s market design.
21 To date, the one market design that has been approved by the SEC that imposes a de minimis delay is that of IEX; the SEC
issued its rules interpretation of “immediate and automatic” simultaneously with its approval of IEX’s exchange application,
with both issued on June 17, 2016. See U.S. Securities and Exchange Commission (2016b) and the related materials referenced
therein. Subsequent to IEX’s approval, the Chicago Stock Exchange (CHX) applied for approval of an asymmetric delay market
design, in which marketable limit orders are slightly delayed, to give liquidity providing quotes a small head start against snipers
in the event of a sniping race. This market design has not been approved. See U.S. Securities and Exchange Commission (2017,
2018a) for the history and public comments regarding CHX’s two versions of the proposal, the latter of which the SEC officially
“stayed” on Oct 24, 2017 and CHX officially withdrew on July 25, 2018. The main substantive argument against the CHX
proposal expressed in public comment letters was that the asymmetry of the delay is inconsistent with the fairness provisions
of the Exchange Act.
22 For further details on intermarket sweep orders see the text of Reg NMS (U.S. Securities and Exchange Commission, 2005).
Formally, the relevant aspects of the regulation are Rule 600(b)(30) for the definition of ISOs, Rule 611(b)(5) for the exchange’s
exemption from ensuring compliance with the Order Protection Rule for ISOs, and Rule 611(c) for this compliance obligation
instead residing in the sender of the ISO.
23 Investors and brokers who do not utilize proprietary data feeds from exchanges instead use a non-proprietary data feed
called the SIP (Securities Information Processor). The SIP feed provides data on the best bid and offer across all exchanges,
10enables market participants to both search available quotes and then “access” them, i.e., trade against them.
First, Rule 610(c) limits the trading fee that any exchange can charge to 0.3 pennies, which, importantly, is
less than the minimum tick size of 1 penny. This ensures that if one exchange has a strictly better displayed
price than another exchange, the price is economically better after accounting for fees. As the SEC writes (pg.
27), “The adopted rule thereby assures order routers that displayed prices are, within a limited range, true
prices.” Second, Rule 610(d) has provisions that together ensure that prices across markets do not become
“locked” or “crossed”—specifically, each exchange is required to monitor data from all other exchanges and
to ensure that it does not display a quote that creates a market that is locked (i.e., bid on one exchange equal
to ask on another exchange) or crossed (i.e., bid on one exchange strictly greater than an ask on another
exchange). Together, then, rules 610(c) and 610(d) ensure that there is a well-defined “national best bid and
offer” (NBBO) across all exchanges (at least ignoring the complexities that arise due to latency, cf. Section 4.
of Budish (2016b)). Third, Rule 610(a) prevents exchanges from charging discriminatory per-share trading
fees based on whether the trader in question does or does not have a direct relationship with the exchange.24
In our model, the notion of a direct relationship with the exchange is captured by the decision of whether
to buy exchange-specific speed technology, which represents exchange products like proprietary data feeds,
co-location, and connectivity. What Rule 610(a) ensures is that market participants face the same trading
fee schedule, whether or not they have such a direct relationship.
To summarize, any time any market participant submits an order, it is required under Rule 611 that
either the market participant themself (if using ISOs) or the exchange they submit their order to checks
quotes on all exchanges. Rule 610 then ensures that this mandatory search is feasible, and that the only
marginal costs of accessing a particular quote on a particular exchange are the exchange’s per-share trading
fees, which are not allowed to be discriminatory. For our theoretical model, therefore, we capture these key
provisions of Reg NMS by assuming what we will call frictionless search and access, on an order-by-order
basis. That is, there is zero marginal cost of search across all exchanges, and there are zero additional
marginal costs (beyond per-share trading fees) of accessing liquidity on a particular exchange or exchanges.
The choice of zero (as opposed to epsilon) is appropriate both because the marginal costs in practice really
are negligible, and because compliance with Rule 611 is mandatory, and zero captures that it is cheaper to
comply with the rule than not to.
3 Theory of the Status Quo
We now develop a simple model of stock exchange competition to better understand the status quo of the
market. The model is a necessary building block for our motivating question about market design innovation,
and is relatively cheap, with fees set by a regulatory process and revenues allocated across exchanges according to a regulatory
formula. However, the SIP feed is slower than proprietary data feeds, primarily because of the time it takes to aggregate
and disseminate data from geographically disparate exchanges. The SIP feed also lacks some additional data that is available
from proprietary feeds, specifically data on depth beyond the best bid and offer, and data on trades of odd lots. One way
to think about the SIP feed is that is appropriate for smaller, non-latency sensitive traders, but not latency-sensitive market
participants. For the purpose of the model, we model the SIP as cheaper (modeled as free) but slower than proprietary feeds.
We discuss exchange revenues from the SIP feed briefly in Section 4.3; we net these revenues out from our estimate of total
exchange-specific speed technology revenues.
24 The prohibition against discriminatory trading fees enables what Reg NMS describes as “private linkages” among exchanges,
which, roughly, are services that provide data about and access to quotes from all exchanges. The text of Reg NMS describes
(pg. 166) that “many different private firms have entered the business of linking with a wide range of trading centers and then
offering their customers access to those trading centers through the private firms’ linkages. Competitive forces determine the
types and costs of these private linkages.” Given our focus on the economics of stock exchanges our model will abstract from
the competition among linkage providers (e.g., broker-dealers, retail brokers) to offer access to end investors; as we discuss in
the conclusion, this seems a fruitful avenue for future research.
11You can also read
